Teenage-AGI — agentic threat model

Utilizes GPT-4 as its foundation model. Primary threats include prompt injection that can hijack the agent's 'thinking' phase, and model alignment risks where the model generates inappropriate or harmful reasoning steps.

Integrates with Pinecone for vector storage to achieve infinite memory recall. This introduces severe risks of memory/knowledge-base poisoning, where malicious inputs are permanently stored and retrieved later, as well as data exfiltration of sensitive personal history stored in the vector database.

Based on a custom Python framework inspired by BabyAGI and Generative Agents. Vulnerable to memory poisoning and state manipulation, where corrupted memory recall disrupts the agent's planning and reasoning loops.

Not certain from the listing — likely run locally as a Python script. Main threats include insecure local storage of OpenAI and Pinecone API keys, and the lack of a sandboxed execution environment for the Python runtime.

Not certain from the listing — no observability, logging, or guardrail mechanisms are mentioned. This creates a significant blind spot, making it difficult to detect when the agent's memory or reasoning has been compromised.

Not certain from the listing — being an open-source hobby/technology project, it lacks enterprise security controls, access management, or compliance alignments (such as GDPR for the 'infinite' personal data stored).

Not certain from the listing — although inspired by interactive simulacra papers, there is no explicit multi-agent coordination or ecosystem integration described, meaning agent-to-agent trust abuse is currently a theoretical risk.